Towards Joint Tardos Decoding: The 'Don Quixote' Algorithm

'Don Quixote' is a new accusation process for Tardos traitor tracing codes which is, as far as we know, the first practical implementation of joint decoding. The first key idea is to iteratively prune the list of potential colluders to keep the computational effort tractable while going from single, to pair,... to t-subset joint decoding. At the same time, we include users accused in previous iterations as side-information to build a more discriminative test. The second idea, coming from the field of mismatched decoders and compound channels, is to use a linear decoder based on the worst case perceived collusion channel. The decoder is tested under two accusation policies: to catch one colluder, or to catch as many colluders as possible. The probability of false positive is controlled thanks to a rare event estimator. We describe a fast implementation supporting millions of users and compare our results with two recent fingerprinting codes.